Detecting scam callers using conversational agent and machine learning systems and methods

Abstract

Systems and methods for detecting indications of a scam caller are disclosed. Call data, such as call audio, is received and used to create a training dataset. Using the training dataset, a machine learning model is trained to detect indications of a scam caller in a phone call. An Interactive Voice Response (IVR) model is trained or configured, using voice samples of speech of a subscriber of a telecommunications service provider, to simulate speech and conversation of the subscriber. A conversational agent is generated using the IVR model and the trained machine learning model. The conversational agent receives a phone call, engages a caller in simulated conversation, and detects indications of whether the caller is a likely scam caller. If the caller is determined to be a likely scam caller, an alert can be generated and/or the call can be disconnected.

I claim: 
     
       1. A computer-implemented method to detect a scam caller using a conversational agent, the method comprising:
 receiving call data for multiple phone calls, 
 wherein at least a portion of the multiple phone calls are associated with known scam callers; 
 creating, using the call data for the multiple phone calls, a training dataset, wherein creating the training dataset includes identifying, in the portion of the multiple phone calls associated with the known scam callers, a set of keywords or phrases indicative of a scam caller, and 
 wherein creating the training dataset includes identifying, in the portion of the multiple phone calls associated with the Known scam callers, at least one topic indicative of a scam caller; 
 training, using the training dataset, a machine learning model to detect indications in one or more received phone calls that an associated caller is a likely scam caller: 
 receiving a set of audio samples associated with a subscriber of a telecommunications service provider; 
 training, based on at least the set of audio samples associated with the subscriber of the telecommunications service provider, an Interactive Voice Response (IVR) model to simulate speech of the subscriber of the telecommunications service provider; 
 generating, using the trained IVR model and the trained machine learning model, a conversational agent to process a phone call from a caller to a mobile device associated with the subscriber of the telecommunications service provider and detect, in the received phone call, indications that the caller is a likely scam caller,
 wherein the conversational agent simulates, in the received phone call, speech of the subscriber of the telecommunications service provider, and 
 wherein the conversational agent analyzes, in the received phone call, speech of the caller to detect indications that the caller is a likely scam caller. 
 
 
     
     
       2. The computer-implemented method of  claim 1 , wherein the received call data for the multiple phone calls includes call audio, and wherein training the IVR model is further based on a second training dataset comprising call audio information of a set of phone calls. 
     
     
       3. The computer-implemented method of  claim 1 , further comprising:
 receiving, by the conversational agent, the phone call from the caller to the subscriber of the telecommunications service provider; 
 detecting, by the conversational agent, indications in the received phone call that the caller is a likely scam caller,
 wherein the indications include at least one of: a keyword, a phrase, or a topic indicative of a scam caller, and 
 wherein detecting the indications includes calculating a confidence score based on the at least one keyword, phrase, or topic indicative of the scam caller; and 
 
 based on detecting the indications that the caller is a likely scam caller, initiating, by the conversation agent, an action comprising at least one of: causing display of a notification on a mobile device associated with the subscriber of the telecommunications service provider, or disconnecting the received phone call. 
 
     
     
       4. The computer-implemented method of  claim 1 :
 wherein the phone call from the caller to the subscriber of the telecommunications service provider is received by the conversational agent in response to determining that the received phone call is likely associated with scam call activity, and 
 wherein determining that the received phone call is likely associated with scam call activity is based, at least in part, on an identifier associated with the caller. 
 
     
     
       5. The computer-implemented method of  claim 1 , wherein the phone call from the caller to the subscriber of the telecommunications service provider is received by the conversational agent in response to determining that the received phone call is likely associated with scam call activity. 
     
     
       6. The computer-implemented method of  claim 1 , further comprising:
 evaluating the trained machine learning model using a testing dataset,
 wherein the testing dataset includes call data associated with known scam callers; and 
 
 when accuracy of the trained machine learning model does not exceed a threshold accuracy, retraining the machine learning model,
 wherein retraining the machine learning model includes at least one of: training the machine learning model at least a second time using the training dataset, resampling at least a portion of the training dataset, or training the machine learning model using a different dataset. 
 
 
     
     
       7. The computer-implemented method of  claim 1 , further comprising:
 evaluating the trained IVR model; and 
 retraining the IVR model using a different set of audio samples associated with the subscriber of the telecommunications service provider, or a different training dataset, or both. 
 
     
     
       8. At least one computer-readable medium, excluding transitory signals, and carrying instructions that, when executed by at least one processor, cause the at least one processor to perform operations comprising:
 receive call data for multiple phone calls, 
 wherein at least a portion of the multiple phone calls are associated with known scam callers; 
 create, using the call data for the multiple phone calls, a training dataset, 
 wherein the at least one processor is caused to calculate at least one variable characterizing the portion of the multiple phone calls associated with known scam callers, and 
 wherein the at least one variable includes a count or frequency of keywords, phrases, pauses, or crosstalk; 
 train, using the training dataset, a machine learning model to detect indications in one or more received phone call that an associated caller is a likely scam caller; 
 receive a set of audio samples associated with speech of a subscriber of a telecommunications service provider; 
 configure, based on at least the set of audio samples associated with speech of the subscriber of the telecommunications service provider, an Interactive Voice Response (IVR) model to simulate speech of the subscriber of the telecommunications service provider; 
 generate, using the IVR model and the trained machine learning model, a conversational agent to process a phone call from a caller to a device 
 associated with the subscriber of the telecommunications service provider and 
 detect, in the received phone call, indications that the caller is a likely scam caller, 
 wherein the conversational agent simulates, in the received phone call, speech of the subscriber of the telecommunications service provider, and 
 wherein the conversational agent analyzes, in the received phone call, speech of the caller to detect indications that the caller is a likely scam caller. 
 
     
     
       9. The at least one computer-readable medium of  claim 8 , wherein the received call data for the multiple phone calls includes call audio, and wherein the operations further comprise:
 train the IVR model based on a second training dataset comprising call audio information of a set of phone calls. 
 
     
     
       10. The at least one computer-readable medium of  claim 8 , wherein the operations further comprise:
 receive, by the conversational agent, the phone call from the caller to the subscriber of the telecommunications service provider; 
 detect, by the conversational agent, indications in the received phone call that the caller is a likely scam caller; and 
 based on detecting the indications that the caller is a likely scam caller, initiate, by the conversational agent, an action comprising at least one of: causing display of a notification on a device associated with the subscriber of the telecommunications service provider, or disconnecting the received phone call. 
 
     
     
       11. The at least one computer-readable medium of  claim 8 :
 wherein the phone call from the caller to the subscriber of the telecommunications service provider is received by the conversational agent in response to determining that the received phone call is likely associated with scam call activity, and 
 wherein determining that the received phone call is likely associated with scam call activity is based, at least in part, on an identifier associated with the caller. 
 
     
     
       12. The computer-readable medium of  claim 8 , wherein the phone call from the caller to the subscriber of the telecommunications service provider is received by the conversational agent in response to determining that the received phone call is likely associated with scam call activity. 
     
     
       13. The computer-readable medium of  claim 8 , wherein the operations further comprise:
 evaluate the trained machine learning model using a testing dataset,
 wherein the testing dataset includes call data associated with known scam callers; and 
 
 when accuracy of the trained machine learning model does not exceed a threshold accuracy, retrain the machine learning model,
 wherein retraining the machine learning model includes at least one of: training the machine learning model at least a second time using the training dataset, resampling at least a portion of the training dataset, or training the machine learning model using a different dataset. 
 
 
     
     
       14. The computer-readable medium of  claim 8 , wherein the operations further comprise:
 evaluate the IVR model; and 
 reconfigure the IVR model using a different set of audio samples associated with the subscriber of the telecommunications service provider, or a different training dataset, or both. 
 
     
     
       15. A computing system, comprising:
 at least one processor; and 
 at least one memory, excluding transitory signals, and carrying instructions that, when executed by the at least one processor, cause the computing system to perform operations comprising: 
 receive call data for multiple phone calls,
 wherein at least a portion of the multiple phone calls are associated with known scam callers; 
 
 create, using the call data for the multiple phone calls, a training dataset,
 wherein the training dataset includes at least one calculated variable based on a count or frequency of a keyword, phrase, pause, or crosstalk included in the portion of the multiple phone calls associated with the known scam callers: 
 
 train, using the training dataset, a machine learning model to detect indications in one or more received phone calls that a caller is a likely scam caller; 
 receive a set of audio samples associated with speech of a subscriber of a telecommunications service provider; 
 configure, based on at least the set of audio samples associated with speech of the subscriber of the telecommunications service provider, an Interactive Voice Response (IVR) model to simulate speech of the subscriber of the telecommunications service provider; 
 generate, using the IVR model and the trained machine learning model, a conversational agent to process a phone call from a caller to a device associated with the subscriber of the telecommunications service provider and detect, in the received phone call, indications that the caller is a likely scam caller, 
 wherein the conversational agent simulates, in the received phone call, speech of the subscriber of the telecommunications service provider, and 
 wherein the conversational agent analyzes, in the received phone call, speech of the caller to detect indications that the caller is a likely scam caller. 
 
     
     
       16. The computing system of  claim 15 , wherein the received call data for the multiple phone calls includes call audio, and wherein the operations further comprise:
 train the IVR model based on a second training dataset comprising call audio information of a set of phone calls. 
 
     
     
       17. The computing system of  claim 15 , wherein the operations further comprise:
 receive, by the conversational agent, the phone call from the caller to the subscriber of the telecommunications service provider; 
 detect, by the conversational agent, indications in the received phone call that the caller is a likely scam caller; and 
 based on detecting the indications that the caller is a likely scam caller, initiate, by the conversational agent, an action comprising at least one of: causing display of a notification on a device associated with the subscriber of the telecommunications service provider, or disconnecting the received phone call. 
 
     
     
       18. The computing system of  claim 15 , wherein the phone call from the caller to the subscriber of the telecommunications service provider is received by the conversational agent in response to determining that the received phone call is likely associated with scam call activity. 
     
     
       19. The computing system of  claim 15 , wherein the operations further comprise:
 evaluate the trained machine learning model using a testing dataset,
 wherein the testing dataset includes call data associated with known scam callers; and 
 
 when accuracy of the trained machine learning model does not exceed a threshold accuracy, retrain the machine learning model,
 wherein retraining the machine learning model includes at least one of: training the machine learning model at least a second time using the training dataset, resampling at least a portion of the training dataset, or training the machine learning model using a different dataset. 
 
 
     
     
       20. The computing system of  claim 15 , wherein the operations further comprise:
 evaluate the IVR model; and 
 reconfigure the IVR model using a different set of audio samples associated with the subscriber of the telecommunications service provider, or a different training dataset, or both.